@gamaze/hicortex 0.4.4 → 0.4.6

package/dist/distiller.js

@@ -205,15 +205,24 @@ async function distillSession(llm, conversation, projectName, date, chunkSizeCha
     }
     // Use provided chunk size or default to no chunking
     const chunkSize = chunkSizeChars ?? MAX_TRANSCRIPT_CHARS;
-    // If transcript fits in one chunk, distill directly
+    // If transcript fits in one chunk, distill directly (errors propagate)
     if (transcript.length <= chunkSize) {
         return distillChunk(llm, transcript, projectName, date);
     }
-    // Chunk large transcripts and distill each segment
+    // Chunk large transcripts and distill each segment.
+    //
+    // Partial success policy:
+    //   - If SOME chunks succeed and SOME fail, return the partial results and
+    //     log a warning. The caller gets *something* and can decide whether
+    //     to count this as success.
+    //   - If ALL chunks fail, throw — no useful output, and the caller needs
+    //     to know this session hit a transient error.
     const chunks = splitIntoChunks(transcript, chunkSize);
     console.log(`[hicortex]     Chunking ${transcript.length} chars into ${chunks.length} segments`);
     const allEntries = [];
     const seen = new Set();
+    let chunkFailures = 0;
+    let lastError = null;
     for (let i = 0; i < chunks.length; i++) {
         console.log(`[hicortex]     Chunk ${i + 1}/${chunks.length} (${chunks[i].length} chars)`);
         try {
@@ -230,30 +239,45 @@ async function distillSession(llm, conversation, projectName, date, chunkSizeCha
         catch (err) {
             const msg = err instanceof Error ? err.message : String(err);
             console.error(`[hicortex]     Chunk ${i + 1} failed: ${msg}`);
-            // Continue with remaining chunks — partial extraction is better than none
+            chunkFailures++;
+            lastError = err instanceof Error ? err : new Error(msg);
         }
     }
+    // If every chunk failed, the session wasn't actually processed. Throw so
+    // the nightly pipeline knows to retry this session next run.
+    if (chunkFailures === chunks.length) {
+        throw lastError ?? new Error("All distillation chunks failed");
+    }
+    if (chunkFailures > 0) {
+        console.warn(`[hicortex]     Partial distillation: ${chunks.length - chunkFailures}/${chunks.length} chunks succeeded`);
+    }
     return allEntries;
 }
 /**
  * Distill a single chunk of conversation text.
+ *
+ * Behaviour contract:
+ *   - Returns `[]` for legitimate empty results (NO_EXTRACT, empty LLM response,
+ *     transcript produced no entries). These are terminal states — the chunk was
+ *     processed successfully, there's just nothing worth keeping.
+ *   - Throws for transient errors (LLM unreachable, HTTP 4xx/5xx, timeout, model
+ *     not found, rate limit). These MUST propagate so the nightly pipeline can
+ *     distinguish "nothing to extract" from "try again later" and avoid
+ *     advancing the last-run watermark past sessions it never actually processed.
  */
 async function distillChunk(llm, transcript, projectName, date) {
     const prompt = (0, prompts_js_1.distillation)(projectName, date, transcript);
-    try {
-        const result = await llm.completeDistill(prompt);
-        if (!result)
-            return [];
-        if (result === "NO_EXTRACT" || result.slice(0, 20).includes("NO_EXTRACT")) {
-            return [];
-        }
-        return parseDistilledEntries(result);
-    }
-    catch (err) {
-        const msg = err instanceof Error ? err.message : String(err);
-        console.error(`[hicortex] Distillation LLM error: ${msg}`);
+    // NOTE: Intentionally no try/catch here. Transient LLM errors (network
+    // failures, 4xx/5xx, model-not-found, timeouts) propagate up to the caller
+    // so the nightly pipeline can treat them as "retry later" instead of
+    // "processed successfully with zero extractions".
+    const result = await llm.completeDistill(prompt);
+    if (!result)
+        return [];
+    if (result === "NO_EXTRACT" || result.slice(0, 20).includes("NO_EXTRACT")) {
         return [];
     }
+    return parseDistilledEntries(result);
 }
 /**
  * Split transcript text into chunks at natural boundaries (double newlines).

package/dist/index.js

@@ -466,28 +466,7 @@ async function autoConfigureLlm(pluginConfig, log) {
         const msg = err instanceof Error ? err.message : String(err);
         log(`[hicortex] LLM test failed (${llmConfig.baseUrl}): ${msg}`);
     }
-    // Step 4: If test failed and this is z.ai, try the alternate endpoint
-    if (llmConfig.provider === "zai") {
-        const altUrl = llmConfig.baseUrl.includes("/coding/")
-            ? llmConfig.baseUrl.replace("/coding/", "/")
-            : llmConfig.baseUrl.replace("/api/paas/", "/api/coding/paas/");
-        log(`[hicortex] Trying alternate z.ai endpoint: ${altUrl}`);
-        llmConfig.baseUrl = altUrl;
-        const altClient = new llm_js_1.LlmClient(llmConfig);
-        try {
-            const response = await altClient.completeFast("Respond with just the word OK", 10);
-            if (response && response.length > 0) {
-                log(`[hicortex] LLM connection verified on alternate endpoint`);
-                persistProviderConfig(llmConfig, log);
-                return llmConfig;
-            }
-        }
-        catch (err) {
-            const msg = err instanceof Error ? err.message : String(err);
-            log(`[hicortex] Alternate z.ai endpoint also failed: ${msg}`);
-        }
-    }
-    // Step 5: Fall back — return the config anyway, log instructions
+    // Step 4: Fall back — return the config anyway, log instructions
     log(`[hicortex] WARNING: Could not verify LLM connection. ` +
         `Distillation and consolidation may fail. ` +
         `To fix: add models.providers.${llmConfig.provider}.baseUrl to ~/.openclaw/openclaw.json ` +

package/dist/init.d.ts

@@ -3,7 +3,7 @@
  *
  * Detection:
  *   1. Local HC server running (localhost:8787)
- *   2. Remote HC server (HICORTEX_SERVER_URL or bedrock:8787)
+ *   2. Remote HC server (HICORTEX_SERVER_URL — any reachable host:port)
  *   3. OC plugin installed (~/.openclaw/openclaw.json)
  *   4. CC MCP already registered (~/.claude/settings.json)
  *   5. Existing DB (~/.hicortex/ or ~/.openclaw/data/)

package/dist/init.js

@@ -4,7 +4,7 @@
  *
  * Detection:
  *   1. Local HC server running (localhost:8787)
- *   2. Remote HC server (HICORTEX_SERVER_URL or bedrock:8787)
+ *   2. Remote HC server (HICORTEX_SERVER_URL — any reachable host:port)
  *   3. OC plugin installed (~/.openclaw/openclaw.json)
  *   4. CC MCP already registered (~/.claude/settings.json)
  *   5. Existing DB (~/.hicortex/ or ~/.openclaw/data/)
@@ -185,10 +185,10 @@ When invoked with \`/learn <text>\`, store the learning in long-term memory via
 
 ## Example
 
-\`/learn z.ai API uses Bearer auth on all endpoints, not x-api-key\`
+\`/learn always check provider docs before assuming an API uses the same auth scheme as OpenAI\`
 
 Becomes a call to hicortex_ingest with:
-- content: "LEARNING: z.ai API uses Bearer auth on all three endpoints (paas, coding, anthropic). Not x-api-key. (2026-03-26)"
+- content: "LEARNING: always check provider docs before assuming an API uses the same auth scheme as OpenAI — header names and token formats vary widely (Bearer vs x-api-key vs custom)."
 - memory_type: "lesson"
 `;
     const learnPath = (0, node_path_1.join)(CC_COMMANDS_DIR, "learn.md");
@@ -402,7 +402,7 @@ async function persistLlmConfig() {
     options.push({
         label: "Other provider (requires API key)",
         save: async () => {
-            console.log("\n  Providers: Anthropic, OpenAI, Google, z.ai, OpenRouter, or any OpenAI-compatible");
+            console.log("\n  Providers: Anthropic, OpenAI, Google, OpenRouter, or any OpenAI-compatible endpoint");
             const baseUrl = await ask("  Provider base URL: ");
             if (!baseUrl) {
                 console.log("  ⚠ Cancelled.");

package/dist/llm.d.ts

@@ -1,6 +1,5 @@
 /**
  * Multi-provider LLM client for consolidation and distillation.
- * Ported from hicortex/consolidate/llm.py.
  *
  * Resolution for OC adapter (resolveLlmConfig):
  *   1. Plugin config (llmBaseUrl, llmApiKey, llmModel)
@@ -10,12 +9,14 @@
  *
  * Resolution for CC adapter (resolveLlmConfigForCC):
  *   1. Explicit env vars (HICORTEX_LLM_BASE_URL + HICORTEX_LLM_API_KEY + HICORTEX_LLM_MODEL)
- *   2. ANTHROPIC_API_KEY → Haiku (cheap, CC users always have this)
- *   3. OPENAI_API_KEY → gpt-4o-mini
- *   4. GOOGLE_API_KEY → gemini-2.0-flash
- *   5. Fallback: Ollama at http://localhost:11434
+ *   2. ANTHROPIC_API_KEY → Claude Haiku (cheap, CC users always have this)
+ *   3. OPENAI_API_KEY → gpt-5.4-nano
+ *   4. GOOGLE_API_KEY → gemini-2.5-flash
+ *   5. Claude CLI fallback (uses subscription, no API key needed)
+ *   6. Fallback: Ollama at http://localhost:11434
  *
- * Supports: OpenAI, Anthropic, Google, OpenRouter, Ollama, z.ai, and 15+ more
+ * Supports any OpenAI-compatible endpoint plus first-class support for
+ * OpenAI, Anthropic, Google, Ollama, OpenRouter, and Claude CLI.
  */
 export interface LlmConfig {
     baseUrl: string;
@@ -68,6 +69,25 @@ export declare function claudeCliConfig(claudePath: string): LlmConfig;
  * Returns the model name if available, null otherwise.
  */
 export declare function probeOllama(baseUrl?: string): Promise<string | null>;
+/**
+ * Pre-flight health check for a specific Ollama endpoint + model.
+ * Returns { ok, reason } so callers can log a clear abort message.
+ *
+ *   - `ok: true` — endpoint reachable AND the requested model appears in
+ *     `/api/tags`. Safe to proceed with a batch distillation run.
+ *   - `ok: false, reason: "unreachable"` — network failure or non-2xx.
+ *   - `ok: false, reason: "model_missing"` — endpoint is up but the
+ *     model isn't listed (the exact case that caused data loss when
+ *     mhac-pro's Ollama didn't have the distill model loaded).
+ *
+ * Matches on exact name OR name prefix ("qwen3.5:35b" matches "qwen3.5:35b-a3b").
+ */
+export declare function probeOllamaModel(baseUrl: string, modelName: string): Promise<{
+    ok: true;
+} | {
+    ok: false;
+    reason: "unreachable" | "model_missing";
+}>;
 /**
  * For batch operations (nightly pipeline), prefer Ollama when available.
  * Claude CLI has strict rate limits that kill batch distillation.
@@ -116,7 +136,7 @@ export declare class LlmClient {
      */
     private completeOllama;
     /**
-     * Anthropic Messages API (/v1/messages). Used for Anthropic and z.ai.
+     * Anthropic Messages API (/v1/messages).
      * Auth via x-api-key header.
      */
     private completeAnthropic;

package/dist/llm.js

@@ -1,7 +1,6 @@
 "use strict";
 /**
  * Multi-provider LLM client for consolidation and distillation.
- * Ported from hicortex/consolidate/llm.py.
  *
  * Resolution for OC adapter (resolveLlmConfig):
  *   1. Plugin config (llmBaseUrl, llmApiKey, llmModel)
@@ -11,12 +10,14 @@
  *
  * Resolution for CC adapter (resolveLlmConfigForCC):
  *   1. Explicit env vars (HICORTEX_LLM_BASE_URL + HICORTEX_LLM_API_KEY + HICORTEX_LLM_MODEL)
- *   2. ANTHROPIC_API_KEY → Haiku (cheap, CC users always have this)
- *   3. OPENAI_API_KEY → gpt-4o-mini
- *   4. GOOGLE_API_KEY → gemini-2.0-flash
- *   5. Fallback: Ollama at http://localhost:11434
+ *   2. ANTHROPIC_API_KEY → Claude Haiku (cheap, CC users always have this)
+ *   3. OPENAI_API_KEY → gpt-5.4-nano
+ *   4. GOOGLE_API_KEY → gemini-2.5-flash
+ *   5. Claude CLI fallback (uses subscription, no API key needed)
+ *   6. Fallback: Ollama at http://localhost:11434
  *
- * Supports: OpenAI, Anthropic, Google, OpenRouter, Ollama, z.ai, and 15+ more
+ * Supports any OpenAI-compatible endpoint plus first-class support for
+ * OpenAI, Anthropic, Google, Ollama, OpenRouter, and Claude CLI.
  */
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.LlmClient = exports.RateLimitError = void 0;
@@ -25,6 +26,7 @@ exports.resolveLlmConfigForCC = resolveLlmConfigForCC;
 exports.findClaudeBinary = findClaudeBinary;
 exports.claudeCliConfig = claudeCliConfig;
 exports.probeOllama = probeOllama;
+exports.probeOllamaModel = probeOllamaModel;
 exports.preferOllamaForBatch = preferOllamaForBatch;
 const node_fs_1 = require("node:fs");
 const node_path_1 = require("node:path");
@@ -76,7 +78,7 @@ function resolveLlmConfigForCC(overrides) {
             baseUrl: overrides.llmBaseUrl,
             apiKey: overrides.llmApiKey,
             model: overrides.llmModel ?? "claude-haiku-4-5-20251001",
-            reflectModel: overrides.reflectModel ?? overrides.llmModel ?? "claude-sonnet-4-5-20250514",
+            reflectModel: overrides.reflectModel ?? overrides.llmModel ?? "claude-sonnet-4-6",
             provider,
         };
     }
@@ -90,7 +92,7 @@ function resolveLlmConfigForCC(overrides) {
             baseUrl: hcBaseUrl,
             apiKey: hcApiKey,
             model: hcModel ?? "claude-haiku-4-5-20251001",
-            reflectModel: process.env.HICORTEX_REFLECT_MODEL ?? hcModel ?? "claude-sonnet-4-5-20250514",
+            reflectModel: process.env.HICORTEX_REFLECT_MODEL ?? hcModel ?? "claude-sonnet-4-6",
             provider,
         };
     }
@@ -101,7 +103,7 @@ function resolveLlmConfigForCC(overrides) {
             baseUrl: process.env.ANTHROPIC_BASE_URL ?? "https://api.anthropic.com",
             apiKey: anthropicKey,
             model: "claude-haiku-4-5-20251001",
-            reflectModel: "claude-sonnet-4-5-20250514",
+            reflectModel: "claude-sonnet-4-6",
             provider: "anthropic",
         };
     }
@@ -111,8 +113,8 @@ function resolveLlmConfigForCC(overrides) {
         return {
             baseUrl,
             apiKey: openaiKey,
-            model: "gpt-4o-mini",
-            reflectModel: "gpt-4o-mini",
+            model: "gpt-5.4-nano",
+            reflectModel: "gpt-5.4-nano",
             provider: detectProvider(baseUrl),
         };
     }
@@ -121,8 +123,8 @@ function resolveLlmConfigForCC(overrides) {
         return {
             baseUrl: "https://generativelanguage.googleapis.com/v1beta",
             apiKey: googleKey,
-            model: "gemini-2.0-flash",
-            reflectModel: "gemini-2.0-flash",
+            model: "gemini-2.5-flash",
+            reflectModel: "gemini-2.5-flash",
             provider: "google",
         };
     }
@@ -150,8 +152,6 @@ function detectProvider(url) {
         return "openrouter";
     if (u.includes("googleapis") || u.includes("generativelanguage"))
         return "google";
-    if (u.includes("z.ai") || u.includes("zai"))
-        return "zai";
     return "openai";
 }
 function readOpenClawConfig() {
@@ -162,7 +162,7 @@ function readOpenClawConfig() {
         const primary = config?.agents?.defaults?.model?.primary;
         if (!primary)
             return null;
-        // primary format is "provider/model" (e.g. "zai/glm-5-turbo", "openai/gpt-4o")
+        // primary format is "provider/model" (e.g. "openai/gpt-5", "anthropic/claude-sonnet-4-6")
         if (typeof primary === "string" && (primary.includes("/") || primary.includes(":"))) {
             const sep = primary.includes("/") ? "/" : ":";
             const [providerHint, ...rest] = primary.split(sep);
@@ -216,7 +216,7 @@ function readOcAuthKey(provider) {
                 const raw = (0, node_fs_1.readFileSync)(authPath, "utf-8");
                 const auth = JSON.parse(raw);
                 const profiles = auth?.profiles ?? {};
-                // Look for a profile matching the provider (e.g. "zai:default")
+                // Look for a profile matching the provider (e.g. "openai:default")
                 for (const [profileId, profile] of Object.entries(profiles)) {
                     const p = profile;
                     if (p?.provider === provider ||
@@ -245,8 +245,8 @@ function resolveFromEnv() {
         return {
             baseUrl,
             apiKey: openaiKey,
-            model: process.env.OPENAI_MODEL ?? "gpt-4o-mini",
-            reflectModel: process.env.OPENAI_MODEL ?? "gpt-4o-mini",
+            model: process.env.OPENAI_MODEL ?? "gpt-5.4-nano",
+            reflectModel: process.env.OPENAI_MODEL ?? "gpt-5.4-nano",
             provider,
         };
     }
@@ -255,8 +255,8 @@ function resolveFromEnv() {
         return {
             baseUrl: process.env.ANTHROPIC_BASE_URL ?? "https://api.anthropic.com",
             apiKey: anthropicKey,
-            model: "claude-sonnet-4-20250514",
-            reflectModel: "claude-sonnet-4-20250514",
+            model: "claude-sonnet-4-6",
+            reflectModel: "claude-sonnet-4-6",
             provider: "anthropic",
         };
     }
@@ -265,8 +265,8 @@ function resolveFromEnv() {
         return {
             baseUrl: "https://generativelanguage.googleapis.com/v1beta",
             apiKey: googleKey,
-            model: "gemini-2.0-flash",
-            reflectModel: "gemini-2.0-flash",
+            model: "gemini-2.5-flash",
+            reflectModel: "gemini-2.5-flash",
             provider: "google",
         };
     }
@@ -280,33 +280,23 @@ function getEnvKeyForProvider(provider) {
             return process.env.ANTHROPIC_API_KEY;
         case "google":
             return process.env.GOOGLE_API_KEY;
-        case "zai":
-            return process.env.ZAI_API_KEY ?? process.env.LLM_API_KEY;
         default:
             return undefined;
     }
 }
-/** Default base URLs for all OC-supported providers (from OC gateway binary). */
+/**
+ * Default base URLs for first-class supported providers.
+ *
+ * For any other provider, set llmBaseUrl explicitly in your config or use
+ * an OpenAI-compatible endpoint. The detectProvider() function will treat
+ * unknown URLs as openai-compatible by default.
+ */
 const PROVIDER_BASE_URLS = {
     openai: "https://api.openai.com/v1",
     anthropic: "https://api.anthropic.com",
     google: "https://generativelanguage.googleapis.com/v1beta",
     ollama: "http://localhost:11434",
     openrouter: "https://openrouter.ai/api",
-    zai: "https://api.z.ai/api/anthropic",
-    groq: "https://api.groq.com/openai/v1",
-    deepseek: "https://api.deepseek.com",
-    mistral: "https://api.mistral.ai/v1",
-    together: "https://api.together.xyz/v1",
-    perplexity: "https://api.perplexity.ai",
-    nvidia: "https://integrate.api.nvidia.com/v1",
-    xai: "https://api.x.ai/v1",
-    venice: "https://api.venice.ai/api/v1",
-    minimax: "https://api.minimaxi.com/v1",
-    moonshot: "https://api.moonshot.ai/v1",
-    kimi: "https://api.kimi.com/coding",
-    chutes: "https://api.chutes.ai",
-    kilo: "https://api.kilo.ai/api/gateway",
 };
 function getDefaultUrlForProvider(provider) {
     return PROVIDER_BASE_URLS[provider.toLowerCase()] ?? "https://api.openai.com/v1";
@@ -378,6 +368,35 @@ async function probeOllama(baseUrl = "http://localhost:11434") {
         return null;
     }
 }
+/**
+ * Pre-flight health check for a specific Ollama endpoint + model.
+ * Returns { ok, reason } so callers can log a clear abort message.
+ *
+ *   - `ok: true` — endpoint reachable AND the requested model appears in
+ *     `/api/tags`. Safe to proceed with a batch distillation run.
+ *   - `ok: false, reason: "unreachable"` — network failure or non-2xx.
+ *   - `ok: false, reason: "model_missing"` — endpoint is up but the
+ *     model isn't listed (the exact case that caused data loss when
+ *     mhac-pro's Ollama didn't have the distill model loaded).
+ *
+ * Matches on exact name OR name prefix ("qwen3.5:35b" matches "qwen3.5:35b-a3b").
+ */
+async function probeOllamaModel(baseUrl, modelName) {
+    try {
+        const resp = await fetch(`${baseUrl.replace(/\/$/, "")}/api/tags`, {
+            signal: AbortSignal.timeout(5000),
+        });
+        if (!resp.ok)
+            return { ok: false, reason: "unreachable" };
+        const data = (await resp.json());
+        const models = data.models ?? [];
+        const found = models.some((m) => m.name === modelName || m.name.startsWith(modelName + ":"));
+        return found ? { ok: true } : { ok: false, reason: "model_missing" };
+    }
+    catch {
+        return { ok: false, reason: "unreachable" };
+    }
+}
 /**
  * For batch operations (nightly pipeline), prefer Ollama when available.
  * Claude CLI has strict rate limits that kill batch distillation.
@@ -505,7 +524,7 @@ class LlmClient {
         if (this.config.provider === "ollama") {
             return this.completeOllama(model, prompt, maxTokens, timeoutMs);
         }
-        if (this.config.provider === "anthropic" || this.config.provider === "zai") {
+        if (this.config.provider === "anthropic") {
             return this.completeAnthropic(model, prompt, maxTokens, timeoutMs);
         }
         return this.completeOpenAiCompat(model, prompt, maxTokens, timeoutMs);
@@ -586,7 +605,7 @@ class LlmClient {
         return result.trim();
     }
     /**
-     * Anthropic Messages API (/v1/messages). Used for Anthropic and z.ai.
+     * Anthropic Messages API (/v1/messages).
      * Auth via x-api-key header.
      */
     async completeAnthropic(model, prompt, maxTokens, timeoutMs) {
@@ -622,8 +641,8 @@ class LlmClient {
      */
     async completeOpenAiCompat(model, prompt, maxTokens, timeoutMs) {
         const baseUrl = this.config.baseUrl.replace(/\/$/, "");
-        // Some providers (z.ai) include version in base URL already
-        const hasVersion = /\/v\d+\/?$/.test(baseUrl) || baseUrl.includes("/paas/v");
+        // Some providers include the API version in the base URL already
+        const hasVersion = /\/v\d+\/?$/.test(baseUrl);
         const url = hasVersion
             ? `${baseUrl}/chat/completions`
             : `${baseUrl}/v1/chat/completions`;
@@ -647,10 +666,6 @@ class LlmClient {
             this.handleRateLimit(resp);
         if (!resp.ok) {
             const text = await resp.text().catch(() => "");
-            // z.ai: "Insufficient balance" likely means wrong endpoint (coding vs paas)
-            if (text.includes("1113") || text.includes("Insufficient balance")) {
-                console.log(`[hicortex] LLM billing error. Check that llmBaseUrl matches your plan. Current: ${baseUrl}`);
-            }
             throw new Error(`LLM API error ${resp.status}: ${text}`);
         }
         const data = (await resp.json());

package/dist/nightly.js

@@ -182,6 +182,24 @@ async function runNightly(options = {}) {
         }
         // Step 2: Distill each session
         let memoriesIngested = 0;
+        let hadTransientFailure = false;
+        // Pre-flight health check for a remote distill endpoint.
+        // If the distill provider is Ollama on a remote host and that host (or the
+        // required model) is unreachable, abort BEFORE touching any sessions —
+        // prevents the data-loss bug where lastRun advances past sessions that
+        // were never actually processed.
+        if (batches.length > 0 && llmConfig.distillBaseUrl && (llmConfig.distillProvider ?? llmConfig.provider) === "ollama") {
+            const distillModel = llmConfig.distillModel ?? llmConfig.model;
+            const health = await (0, llm_js_1.probeOllamaModel)(llmConfig.distillBaseUrl, distillModel);
+            if (!health.ok) {
+                const reason = health.reason === "unreachable"
+                    ? `distill endpoint unreachable (${llmConfig.distillBaseUrl})`
+                    : `distill model not loaded (${distillModel} missing on ${llmConfig.distillBaseUrl})`;
+                console.error(`[hicortex] ABORT: ${reason} — will retry next run, lastRun unchanged`);
+                hadTransientFailure = true;
+                batches.length = 0; // Skip the distillation loop entirely
+            }
+        }
         // Detect safe chunk size based on model context window
         const chunkSize = await (0, distiller_js_1.detectChunkSize)(llmConfig.provider, llmConfig.distillModel ?? llmConfig.model, llmConfig.baseUrl);
         for (const batch of batches) {
@@ -190,6 +208,20 @@ async function runNightly(options = {}) {
                 console.log(`[hicortex]   Skip ${batch.sessionId.slice(0, 8)} (${batch.projectName}): too short`);
                 continue;
             }
+            // Server-mode per-session dedup: skip sessions already in the DB.
+            // Client mode gets this for free via the server's /ingest endpoint;
+            // server mode writes directly via storage.insertMemory and needs
+            // an explicit check. This makes retries of previously-failed runs
+            // idempotent.
+            if (!dryRun) {
+                const existing = db
+                    .prepare("SELECT COUNT(*) as c FROM memories WHERE source_session = ?")
+                    .get(batch.sessionId);
+                if (existing.c > 0) {
+                    console.log(`[hicortex]   Skip ${batch.sessionId.slice(0, 8)} (${batch.projectName}): already ingested`);
+                    continue;
+                }
+            }
             console.log(`[hicortex]   Distilling ${batch.sessionId.slice(0, 8)} (${batch.projectName}, ${batch.date})`);
             if (dryRun) {
                 console.log(`[hicortex]     [dry-run] Would distill ${transcript.length} chars`);
@@ -224,7 +256,8 @@ async function runNightly(options = {}) {
             }
             catch (err) {
                 const msg = err instanceof Error ? err.message : String(err);
-                console.error(`[hicortex]     Distillation failed: ${msg}`);
+                console.error(`[hicortex]     Distillation failed: ${msg} — will retry next run`);
+                hadTransientFailure = true;
             }
         }
         console.log(`[hicortex] Distillation complete: ${memoriesIngested} new memories`);
@@ -241,8 +274,17 @@ async function runNightly(options = {}) {
             console.log(`[hicortex] CLAUDE.md updated: ${injection.lessonsCount} lessons at ${injection.path}`);
         }
         // Step 5: Update last-run timestamp
+        // CRITICAL: only advance lastRun if every session was processed without
+        // a transient failure. Otherwise failed sessions would be permanently
+        // lost — they'd be older than the new lastRun and never retried.
         if (!dryRun) {
-            writeLastRun();
+            if (hadTransientFailure) {
+                console.warn(`[hicortex] Not advancing lastRun — one or more sessions failed. ` +
+                    `They will be retried on the next run.`);
+            }
+            else {
+                writeLastRun();
+            }
         }
         console.log(`[hicortex] Nightly pipeline complete.`);
     }
@@ -329,6 +371,22 @@ async function runClientNightly(config, dryRun) {
             writeLastRun();
         return;
     }
+    // Pre-flight health check for a remote distill endpoint (client mode).
+    // If the distill provider is Ollama on a remote host and the required model
+    // isn't loaded, abort BEFORE touching any sessions — same data-loss fix
+    // as server mode.
+    let hadTransientFailure = false;
+    if (llmConfig.distillBaseUrl && (llmConfig.distillProvider ?? llmConfig.provider) === "ollama") {
+        const distillModel = llmConfig.distillModel ?? llmConfig.model;
+        const health = await (0, llm_js_1.probeOllamaModel)(llmConfig.distillBaseUrl, distillModel);
+        if (!health.ok) {
+            const reason = health.reason === "unreachable"
+                ? `distill endpoint unreachable (${llmConfig.distillBaseUrl})`
+                : `distill model not loaded (${distillModel} missing on ${llmConfig.distillBaseUrl})`;
+            console.error(`[hicortex] ABORT: ${reason} — will retry next run, lastRun unchanged`);
+            return; // Don't touch lastRun; next trigger retries the same sessions
+        }
+    }
     // Distill each session and POST to server
     let memoriesIngested = 0;
     let sessionsSent = 0;
@@ -388,6 +446,7 @@ async function runClientNightly(config, dryRun) {
                 }
                 else {
                     console.error(`[hicortex]     Ingest failed (${resp.status}): ${result.error}`);
+                    hadTransientFailure = true;
                 }
             }
             if (sessionCount > 0) {
@@ -396,7 +455,8 @@ async function runClientNightly(config, dryRun) {
             }
         }
         catch (err) {
-            console.error(`[hicortex]     Failed: ${err instanceof Error ? err.message : String(err)}`);
+            console.error(`[hicortex]     Distillation failed: ${err instanceof Error ? err.message : String(err)} — will retry next run`);
+            hadTransientFailure = true;
         }
     }
     // Inject lessons from server into CLAUDE.md
@@ -408,8 +468,17 @@ async function runClientNightly(config, dryRun) {
             console.error(`[hicortex] CLAUDE.md injection failed: ${err instanceof Error ? err.message : String(err)}`);
         }
     }
-    if (!dryRun)
-        writeLastRun();
+    // Only advance lastRun if every session was processed without a transient
+    // failure. Otherwise failed sessions would be permanently lost.
+    if (!dryRun) {
+        if (hadTransientFailure) {
+            console.warn(`[hicortex] Not advancing lastRun — one or more sessions failed. ` +
+                `They will be retried on the next run.`);
+        }
+        else {
+            writeLastRun();
+        }
+    }
     console.log(`[hicortex] Client nightly complete: ${memoriesIngested} memories from ${sessionsSent} sessions → ${serverUrl}`);
 }
 /**

package/openclaw.plugin.json

@@ -2,7 +2,7 @@
   "id": "hicortex",
   "name": "Hicortex — Long-term Memory That Learns",
   "description": "Your agents remember past decisions, avoid repeated mistakes, and get smarter every day. Nightly reflection generates actionable lessons that automatically update agent behavior.",
-  "version": "0.3.11",
+  "version": "0.4.6",
   "kind": "lifecycle",
   "skills": ["./skills/hicortex-memory", "./skills/hicortex-learn", "./skills/hicortex-activate"],
   "configSchema": {
@@ -10,7 +10,7 @@
     "properties": {
       "licenseKey": {
         "type": "string",
-        "description": "Hicortex license key (hctx-...). Leave empty for free tier (100 memory cap)."
+        "description": "Hicortex license key (hctx-...). Leave empty for free tier (250 memory cap)."
       },
       "llmBaseUrl": {
         "type": "string",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gamaze/hicortex",
-  "version": "0.4.4",
+  "version": "0.4.6",
   "description": "Human-like memory for self-improving AI agents. Automatic capturing, nightly reflection, and cross-agent learning. Works with Claude Code and OpenClaw.",
   "main": "dist/index.js",
   "bin": {

package/skills/hicortex-learn/SKILL.md

@@ -23,12 +23,12 @@ When invoked with `/learn <text>`, store the learning in long-term memory via hi
 ## Example
 
 ```
-/learn z.ai API uses Bearer auth on all endpoints, not x-api-key
+/learn always check provider docs before assuming an API uses the same auth scheme as OpenAI
 ```
 
 Becomes:
 ```
-hicortex_ingest(content="LEARNING: z.ai API uses Bearer auth on all three endpoints (paas, coding, anthropic). Not x-api-key. (2026-03-24)", project="global", memory_type="lesson")
+hicortex_ingest(content="LEARNING: always check provider docs before assuming an API uses the same auth scheme as OpenAI — header names and token formats vary widely (Bearer vs x-api-key vs custom). (2026-04-07)", project="global", memory_type="lesson")
 ```
 
 ## Rules